# ZJX 1.26 RELICWIRE public source note Published: 2026-09-04 Verified: 2026-09-04 Status: latest promoted ZJX version; public distribution not announced ## Claim boundary ZJX 1.26 RELICWIRE recovered a conditional one-pass fast path from the ZJX 1.24 GHOSTLANE campaign, which had closed as a terminal failure. The recovered candidate was ported onto the released ZJX 1.25.1 NIGHTMARKET reference, measured against frozen thresholds on a quiesced measurement host, audited internally, repaired, and reproduced on a second clean boot before promotion. ZJX 1.26 RELICWIRE is the latest promoted ZJX version. ZJX 1.25.1 NIGHTMARKET is the previous public release. Public distribution of 1.26 is a separate decision and has not been announced. Update, 2026-09-04: the product site now leads with the promoted 1.26 version. The original edition of this note kept 1.25.1 as the current public distribution while reporting the 1.26 promotion. This update separates product promotion from binary availability; it does not announce a public download. This source note publishes the disposition, the bounded benefit figures, the correctness outcomes, the audit disposition, and the limitations needed to evaluate the public claims. The full Chronicle contains the audit findings, the restoration repair, the reproduction protocol, the size-eligibility analysis, and the disclosed limitations, and requires an unlock. Canonical Chronicle: Previous public release: Recovered experiment: Product page: ## Mechanism identity Before compressing a large solid stream, the encoder asks its resource governor a question that mutates no state: if the complete worst-case output buffer were required now, could it be admitted? When admission succeeds the encoder takes a single compression pass. When admission is refused it runs the inherited two-pass path unchanged. Archives are byte-identical on both paths, so the choice is invisible to the archive format, the decoder, and every consumer. A genuine compression error never falls back to the other path. No format, decoder, candidate-universe, candidate-order, scoring, schedule, threshold, metadata, or command-line change accompanies the mechanism. ## Bounded result Measured on the isolated measurement host, paired and seeded, against the released reference, on the compact objective: - large single-file text corpus: about 28 percent lower wall time and process CPU, three of three favourable comparisons - large mixed public corpus: about 43 percent lower wall time and process CPU, three of three favourable comparisons - peak memory multiplier: about 1.00 against a frozen allowance of 1.10 - archive byte identity: all six measured pairs in the repaired second-boot reproduction - exact restoration: all twelve measured archives in the repaired second-boot reproduction These are corpus-specific results on large solid workloads under authorized commands and frozen thresholds. They are not a universal ZJX speed claim and not a claim of compressor dominance. ## Size eligibility The changed code is reachable only for solid streams inside a bounded size band. Below the lower bound the encoder takes a different route and the mechanism cannot execute. Small-corpus behaviour is therefore provably unchanged rather than merely unmeasured. ## Audit disposition Before promotion the campaign was audited internally under instructions to treat every conclusion as an unverified claim. The audit returned two critical findings and refused to clear the release: 1. No restoration evidence existed. Every measured run recorded a null restored-tree digest because the harness carried a hard-coded placeholder in place of the unpack-and-compare step, so the check compared null against null and passed unconditionally. 2. The promotion label was not permitted. The governing specification requires an independent reproduction on a second clean boot; the reproduction had been attempted on the same boot, had never produced a passing verdict, and had been recorded as partial. Neither finding was a defect in the shipped code. Both were defects in the evidence and in what the evidence had been claimed to show. ## Repair and reproduction Restoration verification was rebuilt to fail closed. Every measured archive is now inspected, tested, unpacked, and compared against a canonical digest of the source tree including permissions and nanosecond timestamps. Missing, added, renamed, corrupted, truncated, or metadata-divergent restorations all move the digest, and a run without a matching pair carries no verdict. The independent reproduction ran on a boot distinct from the original campaign, with fresh source extraction, fresh builds under the pinned toolchain, an executor independent of the campaign, a quiesced host, and opening and closing control blocks. All six frozen gates passed, within half a percentage point of the original campaign on both corpora and both metrics. Twenty measured runs produced zero environment breaches, and restoration was exact in all twelve measured archives. ## Limitations - Benefit figures are bounded to one measurement host, two corpora, authorized commands, and frozen thresholds. - The reproduction runner was not blind to the expected aggregates. The specification prefers a blind runner. - One non-target ceiling on a fast-objective workload remains corroborated rather than formally decided. Every measurement of it is favourable or inside noise and none shows a regression, but it is not a verdict. - Repository identity, object identifiers, host paths, binary digests, internal function and fixture names, and raw traces are excluded from this published note.